Image: GitHub
Why it mattersEvery skill a coding agent loads sits in the context window on every turn. A rewrite that keeps the skill count but shrinks the per-turn cost is the right trade for a team.
Every agent skill a team installs costs tokens on every turn. Phuryn released product-engineering on 29 October, a free MIT bundle of 43 skills for releasing with coding agents. The author rewrote the skill descriptions and the per-turn load fell by about three quarters.
The repository is attributed to The Product Compass. It has 152 stars three days after it was created, with 23 forks and 53 commits on main. The bundle names six AI coding assistants it supports: Claude Code, Codex, Cursor, Gemini CLI, OpenCode and GitHub Copilot. The README says it works with "any other agent that reads SKILL.md folders".
What is in it
The 43 skills sit in six packs. The product pack covers deciding what to build: start-here, propose, competitor-analysis, market-segments, beachhead-segment, value-proposition, value-curve and growth-loops. The ai-shipping pack covers getting from code to release: ship-check, code-review, intended-vs-implemented, shipping-artifacts, release-gate, test-harness, document-app, derive-tests, security-audit-static and performance-audit-static. The other four packs cover agent teams and handovers. They also cover app stack choices like Supabase and Clerk, and a text editor that rewrites AI-sounding prose.
The README says 27 of the 43 skills were learned on real products built with Claude Code, Codex and Grok. Each of the 27 keeps dated notes on when it broke or was measured. The other 16 skills do not carry that history.
The token-budget rewrite
The README states a specific per-turn cost. Loading every pack in Claude Code adds about 2,600 tokens to every turn. The number was about 9,100 before the author rewrote the descriptions. The 39 self-starting skills share a 6,000-character total, with 300 characters each. A validator written for the bundle enforces the limit, so a new skill cannot push the budget back up.
What that number means for a reader. An agent skill in this framework sits in the system prompt on every call. The agent sees the short description of each skill on every turn. It reads the full skill file only when it decides to use one. Rewriting the descriptions keeps the full set available at a lower per-turn cost. A team can copy the method with or without installing the bundle. Setting a 300-character limit per skill is the move. The limit does the work.
The honest read on the number is that the author ran it on their own install. A different team's set of skills, a different agent harness, or a different project context may land somewhere else. The direction of the saving is the readable part. The lesson a reader can act on is the character budget per skill, not the exact 2,600.
Source
product-engineering on GitHub, The Product Compass.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.